Be honest about the answer. Amazon’s ranking systems allocate demand across millions of sellers. Uber’s dispatch and surge algorithms plan a city’s transport supply in real time. Google and Meta run auctions, but the auctioneer sets the rules, the reserve prices, and the measurement of success. App stores decide which software businesses exist. Credit scores decide which households are liquid, and predictive policing systems — bought from vendors, procurement by procurement, in exactly the pattern O’Neil’s catalogue documents — decide where the state looks. None of this is a market in the sense your economics textbook meant. It is allocation by central authority, executed at a scale Gosplan’s economists would have wept over.

We spent the twentieth century proving that central planning fails. Then we rebuilt it in the private sector and called it a platform.

The debate we think we won

Quick history, because the framework pays rent. In 1920, Ludwig von Mises argued that a socialist economy literally could not calculate: without private property there are no exchanges, without exchanges no prices, and without prices no way to compare the value of ten thousand alternative uses for steel. Friedrich Hayek sharpened the point in 1945 — the knowledge needed for allocation exists as millions of dispersed, local, tacit fragments, and the price system is the only known machine that aggregates them. The socialist side, led by Oskar Lange, replied that planners could simulate prices and adjust by trial and error. Seventy years of shortage economies settled the argument. Markets: 1. Plans: 0.

James C. Scott’s Seeing Like a State told the same story anthropologically: the planner’s map is thinner than the territory, and ruling through the map destroys what the map can’t see. So far, this is the comfortable version every business reader knows.

The uncomfortable version: Mises’s argument was about mechanism, not ownership. Nothing in it says planning becomes rational when the planner is incorporated in Delaware. If allocation without price signals fails, it fails wherever the prices stop — and inside the walls of a platform, the prices stop. The seller doesn’t bid for shelf position in an open market; a ranking model assigns it. The driver doesn’t negotiate a fare; a dispatch system posts one, take it or leave it. Ronald Coase told us in 1937 that firms are islands of conscious planning in the market sea. Fine — but an island the size of a continent, intermediating the commerce of whole economies, is not what he had in mind. The calculation debate isn’t over. It moved.

The market test did move up a level — that much of the steelman holds. So when does the higher-level test stop disciplining the planner? Three conditions, and you can check each one against a balance sheet.

Network effects turn exit into a collective action problem. A merchant can leave Amazon the way a Soviet factory manager could refuse a quota — technically, once. When the customers, the reviews, the fulfillment, and the search demand all live inside the venue, individual exit is unilateral disarmament. Discipline requires that defection be individually rational; network effects make it individually ruinous. That’s the moat, and the moat’s whole job is to blunt the market test.

Switching costs convert users into assets. Watch where the margin comes from. When a platform’s take rate rises year over year while its service quality plateaus — advertising loads up in the search results, the algorithm favors house brands, dispatch pay tightens — you are watching a planner harvest locked-in participants, not a competitor winning contested business. Cory Doctorow coined the word for this trajectory — enshittification — and the mechanism under the coinage is textbook: once acquiring a replacement platform costs participants more than the deteriorating terms, terms deteriorate. Prices inside the venue no longer carry information; they carry rent.

Regulatory capture closes the loop. The final discipline on a dominant planner is the state, and the planner knows it. Compliance regimes priced for incumbents, licensing moats, procurement relationships — the same playbook every administrative monopoly has run since railroads. When the referee’s rulebook is co-authored by the biggest player, the “higher-level market test” has been acquired like any other supplier.

Where the three conditions are absent — contested markets, multi-homing users (participants who operate across rival venues at once), credible entrants — platform planning stays disciplined and mostly beneficial. Where they’re present, you get exactly the pathologies the calculation debate predicted, with better UX: allocation drifting toward what the planner can measure and monetize, information flowing up the hierarchy getting gamed (five-star inflation is quota-report inflation), peripheral participants over-investing in legibility to the center — in being easily read and rated by the platform — instead of in value to the customer. O’Neil’s scored teachers and Scott’s schematic forests were the micro and the macro of the same drift; the platform economy is its middle layer. And Kate Crawford’s Atlas of AI traces the appetite down the physical stack — labeled data, minerals, logistics, click-work — in case anyone still thinks the dashboard is the whole machine rather than the visible end of a planning empire’s supply chain.

The series, closed

This is part three of an argument. Part one showed software agents defecting to trivial transactions because disposable identities kill the shadow of the future. Part two showed scoring systems doing the same thing to humans — unappealable, asymmetric, final. Here’s the institutional layer: whoever owns the venue sets the length of everyone’s shadow. A platform that can deactivate you without appeal has made your future short. A planner disciplined by exit and competition has to keep your future long enough that you’ll invest in it. The stag hunt, the score, and the plan are one problem at three scales, and the variable is always who controls the horizon.

The multi-homing test

So what? Three decisions.

If you operate on platforms: measure your dependence like counterparty risk. What fraction of revenue arrives through venues you can’t exit within a quarter? That number is the share of your business plan that is actually someone else’s plan.

If you build platforms: the moat math has a second column now. Rents extracted from locked-in participants show up later as regulatory risk, seller flight to entrants, and political salience. The durable position — and part one of this series made the same bet for agent markets — is owning the venue participants choose while multi-homing, because trust is the one input a captive market can’t fake.

If you regulate platforms: skip the ideology and audit one variable — can participants multi-home? Where sellers, drivers, and developers realistically operate across venues, competition is doing the disciplining and intervention should be light. Where they can’t, you are looking at a planning bureau with shareholders, and the relevant policy tradition is not antitrust-as-consumer-prices but common carriage, interoperability mandates, and appeal rights — the tools every society eventually applies to unaccountable allocators.

And a marker for falsification: watch take rates, self-preferencing, and participant margins at the platforms with the strongest lock-in. If over the coming years the captive venues behave no differently from the contested ones, the calculation-debate framing failed its own audit — cut it from the deck. The Soviet planners had no such test. That we still do is the remaining difference worth defending.

The lesson has only one defect, which is that its central premise has quietly become false. It is the purpose of this essay to say so plainly, and then to rescue what remains — which is, as it happens, the more important half.

What Scott actually claimed

Scott’s argument deserves to be stated carefully, since it is usually invoked rather than read. States, he observed, can only govern what they can see, and so they remake the world to be seeable: standard surnames, uniform land registers, monoculture forests, cities on grids. The instrument of this remaking is the schema — a simplified formal representation that keeps the features useful to power and discards the rest. What the schema discards Scott called metis: the practical, adaptive, largely unwritable knowledge of practitioners — how this river floods, when this soil tires, which neighbor’s promise is good.

Half a century before Scott’s cadastral maps, the same argument had appeared in economic dress. Friedrich Hayek’s 1945 essay “The Use of Knowledge in Society” argued that the knowledge that matters to an economy exists as dispersed, tacit, local fragments — “knowledge of the particular circumstances of time and place” — which no central authority can assemble, because the assembling destroys it. Scott is Hayek’s knowledge problem in anthropological costume; that the two men would have disagreed about nearly everything else is one of the small ironies that make the history of ideas worth reading. Both rest on a single load-bearing premise: tacit knowledge cannot be centralized. You may employ the man who knows the harbor, but you cannot extract the harbor from the man.

For two centuries of administrative technology, the premise held. Every instrument the centralizers built — the census, the survey, the standard form, the relational database — could store only what could first be made explicit. The schema came first; whatever fit it, survived; whatever did not, vanished. On this premise the entire critical tradition rests.

The premise fails

Now consider what a large language model is. It is not a schema that the messy world was forced into. It is a statistical compression of the mess itself — trillions of words of working notes, forum answers, code reviews, kitchen improvisations, harbor lore — written by millions of practitioners for one another, in their own unstandardized idiom. Nobody designed its categories, and nobody could: its “categories” are whatever regularities the text of human practice happens to contain.

The results are familiar to anyone who has used the technology honestly. Ask a model why a sourdough starter behaves differently in a humid kitchen, how to calm a particular kind of anxious customer, what an error message usually means as opposed to what it says — and you receive something that behaves remarkably like metis: context-sensitive, hedged, alive to the exceptions. It follows that the first technology in history has appeared which ingests tacit knowledge without first making it explicit. The extraction that Hayek declared impossible and Scott declared destructive is now performed, imperfectly but at planetary scale, by gradient descent.

To this the traditional critic replies that the model does not really understand the harbor. Perhaps not; the question is interesting and this essay will not settle it. But the reply misses the practical point. Scott’s argument never depended on the state’s comprehension — the Prussian forestry office understood nothing; the damage came from what its instruments could and could not carry. The new instruments carry the mess. A critique aimed at the limitations of the carrier must reckon with a carrier that no longer has those limitations, and the honest conclusion is uncomfortable: the naive claim that “AI destroys context” is dead, and its defenders should stop making it. An argument kept alive after its premise has died persuades no one but its owners, and discredits the sounder arguments standing next to it.

The relocation

Where, then, did the legibility problem go? It did not dissolve; it changed address — from the representation to the supply chain — and Kate Crawford’s Atlas of AI is, in effect, its new cadastral survey.

Consider what must be made legible in order that the model may remain illiterate in the old, blessed sense. The workers who label training data are managed through piecework platforms that score them in fractions of a cent, their judgments forced into taxonomies precisely as rigid as any Prussian forest register. The text of human practice is scraped under terms no practitioner negotiated, catalogued, deduplicated, filtered — an enclosure of the written commons conducted with the same serene confidence as the enclosures of land that gave Scott’s states their fields. The minerals are mined, the electricity metered, the data centers sited, each through administrative schemata of the classical kind. And when the model’s fluent output is bolted into a bureaucracy — a benefits office, a content moderation pipeline, a hiring funnel — it is wrapped in exactly the scores, thresholds, and appeal-proof categories that this site’s other essays examine. The mess is in the middle; the grid closes over both ends.

It follows that Scott requires not refutation but a change of address. The violence of simplification, which the critical tradition kept looking for inside the representation, now operates in the industrial process that produces the representation and in the institutional process that consumes it. One may put the revised thesis in a single sentence: deep learning did not abolish the legibility problem; it relocated the problem from the map to the map-making economy.

Nor has the claim shrunk in the move; it has arguably grown, for the old schemas were at least public — one could read the census categories and object to them. A supply chain is private, distributed across jurisdictions, and visible chiefly to those who own it. The cadastral map has, so to speak, gone underground, and critics who continue to inspect the surface will report, quite accurately, that they cannot find it.

What remains of the old religion

The reader may reasonably ask what practical difference the relocation makes. Three follow at once.

First, the target of reform changes. If the trouble were in the representation, the remedy would be better representations — richer categories, “context-aware” models — and the industry would be delighted to sell them to us. If the trouble is in the supply chain, the remedies are labor law for labelers, provenance and consent for training corpora, and public accounting of material costs: dull instruments, which is generally the mark of real ones.

Second, the evaluative question changes. One should ask of any AI deployment not “does the model understand the context?” — it increasingly does, or counterfeits understanding to a standard the question cannot detect — but “what had to be gridded so that this fluency could be delivered here?” The answer is an inventory of labelers, scrapes, mines, and scores, and it differs from deployment to deployment in ways that permit actual judgment.

Third, a prediction, since a thesis that risks nothing is worth nothing. If the relocation view is right, the harms that surface over the coming years will cluster not around models flattening meaning — the translation will keep getting better — but around the two gridded ends: disputes over what was taken to train, and injuries done by the institutional casings into which the fluent middle is fitted. If instead the characteristic scandals of the next decade are semantic — machines catastrophically misreading context in ways their supply chains cannot explain — then Scott’s original address was correct after all, and the relocation argued here will have been a detour: an error one could hardly regret discovering, since it would mean the maps were once again lying where everyone could inspect them.

The intelligent position, in short, is neither the triumphalism that declares the knowledge problem solved nor the nostalgia that repeats a critique whose premise has expired. The state has finally learned to read our handwriting. The urgent question is what it costs everyone downstream of the reading lesson — and that question belongs to Scott still, provided his heirs consent to follow the problem to its new address.

This matters now because building fast is back on the agenda, and this time the builders are right. Ezra Klein and Derek Thompson’s Abundance argues that the rich world’s housing shortages, energy constraints, and infrastructure paralysis are self-inflicted — procedural sclerosis, not physical limits. The diagnosis is correct. The prescription — build faster, permit faster, say yes more — is directionally correct. And it is exactly the mood in which every disaster in Scott’s catalogue was launched.

The engineering question, then, is not whether to pursue abundance but how to pursue it without rebuilding the failure modes. That question has an operational answer, and it comes from an unexpected direction: the practices that software and manufacturing developed after their high-modernist phase failed.

Two kinds of knowledge, one missing from the blueprint

Scott’s central term needs a precise definition, because the whole argument turns on it. Metis (from the Greek for cunning intelligence) is practical, local, adaptive knowledge — the kind a harbor pilot has and a hydrodynamics textbook does not. It is learned by doing, held by practitioners, and largely impossible to write down. Its opposite number is the schematic, standardized, transmissible knowledge a central planner works with: maps, models, codes, averages.

The high-modernist failure pattern is structural, and it repeats:

• Compression. The plan reduces a territory to the variables the planner can see. The German foresters saw board-feet; they did not see the soil fungi, insect ecologies, and undergrowth that kept the forest alive. First harvest: excellent. Second: collapse, and a new German word — Waldsterben, forest death.
• Amputation. Implementation deletes the local practice that was silently doing the work. Brasília’s planners eliminated the street; they got a capital famous for having nowhere to be a pedestrian, ringed by unplanned satellite cities where the actual life happens.
• Lock-in. The plan executes at full scale before feedback can arrive, so by the time the error signal shows up, it is the size of the country. Collectivization’s feedback arrived as famine.

The pattern needs no villains and no bad taste to run — only systems built with no error-correction path: high coupling, no rollback, feedback latency measured in years. Any engineer who has shipped a big-bang rewrite recognizes the shape immediately.

The toolkit that already solved this

Software’s own high-modernist era — waterfall planning, multi-year specifications, integration at the end — died of the same disease, and the practices that replaced it are a working answer to Scott’s critique, not a metaphor for one:

• Small batches. Ship the smallest increment that produces a real signal. Batch size is the master variable: it bounds the blast radius of every wrong assumption.
• Safe-to-fail probes. In domains where cause and effect are only clear in hindsight — what the Cynefin framework calls complex domains — you do not analyze your way to the answer. You run cheap parallel experiments, amplify what works, and kill what doesn’t. The probe is designed to be survivable before it is designed to succeed.
• Fast feedback, tight loops. The value of feedback decays with latency. A code review in an hour changes the design; the same finding a year later is archaeology.
• Reversibility as a requirement. Feature flags, canary deployments, rollback plans: modern infrastructure treats “how do we undo this?” as a design input, not an incident-response question.

Translate the toolkit to the physical world and you get a recognizable abundance politics, just not a Promethean one: zoning reform that legalizes the next increment of density everywhere (small batches) rather than master-planning megaprojects; dozens of parallel geothermal and storage pilots (probes) rather than one bet-the-decade facility; permitting that returns an answer in ninety days (loop latency) as the reform metric, rather than permitting that returns a perfect answer never.

None of this counsels timidity — it is how you go fast without betting the system on your model of it being right. The opposite of stagnation is not centralization. It is feedback.

The steelman contains two distinct objections, and they deserve separate treatment.

The lumpy-asset objection is real physics, and the answer is to move the iteration, not to deny the lump. A tunnel is atomic; a tunnel program is not. France built fifty-six reactors in two decades not by treating each as a bespoke experiment but by standardizing a handful of designs and letting each unit feed corrections into the next — iteration at the fleet level, wrapped around irreducibly lumpy units. The US built reactors as one-off artisanal projects and got cost curves that bent the wrong way. Same physics, opposite feedback architecture. Where even the program is a single shot — one harbor barrage, one megadam — the honest engineering position is that these are the projects that earn high-modernist rigor: exhaustive reference-class forecasting, independent red teams, and the humility to treat “reversible” as meaning “we wrote down the decommissioning plan and funded it,” because for a nuclear plant that is what the word can mean. Incrementalism hits a boundary at physics; the discipline of bounded bets does not.

The vetocracy objection mistakes veto latency for feedback. A feedback loop measures the world’s response to the thing you built. An environmental review that takes seven years before anything exists is not a feedback loop; it is open-loop delay — worse than none, because it consumes the error budget before the first real signal arrives. The distinction is operational: feedback is fast, empirical, and consequence-bearing (the pilot plant underperforms, the next one changes); vetoes are slow, predictive, and consequence-free for the vetoer. A regime of “build the increment, measure it, let the measurement gate the next increment” is more permissive than the status quo precisely because it moves judgment from forecasts to results. Klein and Thompson want the state to deliver; delivery is a control loop, and control loops need plants that actually run.

What this costs

There are no solutions, only trade-offs — Sowell’s rule — so state them. A feedback-first abundance agenda gives up the outsized wins of a moonshot in the cases where the moonshot would have worked. It accepts coordination overhead: many probes need portfolio management that one megaproject does not. It requires institutions capable of killing their own pilots, which is politically the hardest kill in government. And it will sometimes be slower to start — though rarely slower to finish, because the megaproject’s schedule was fiction all along.

What it buys is survivability of error, and that is the correct thing to optimize, because the defining property of building in complex domains is that some of your confident assumptions are wrong and you do not know which.

A falsifiable version of the claim, for the record: across infrastructure categories, delivery programs with shorter build-measure cycles and standardized repeatable units should show flatter real-cost curves over time than one-off master-planned equivalents — nuclear France versus nuclear America is one existing data point, and the comparative transit-costs literature already documents the Madrid-versus-New-York version of the same gap. Assemble the dataset; if the correlation runs the other way, the feedback-first position stops being an engineering claim and becomes a taste, and should be weighted accordingly.

Scott ends Seeing Like a State not against planning but against imperial planning — the kind that silences the feedback of the governed and the ground. The abundance movement is young enough to choose its ancestor. It can descend from the moonshot, and re-run the century’s most expensive experiment in ignoring the terrain. Or it can descend from the practices that learned, at industrial scale, how to let the terrain talk back — and build more, faster, because the error signals arrive while they are still cheap.

This essay is the second part of an argument that began with The High-Speed Stag Hunt, which examined why autonomous software agents default to low-value, risk-free transactions. The mechanism there was game-theoretic: when identities are disposable and memory is short, cooperation stops paying. This time the subject shifts from agents negotiating with agents to people living inside scoring systems. The claim is that the same mechanism operates on them — from the opposite side of the interface.

A term worth defining precisely

Robert Axelrod, whose 1984 book The Evolution of Cooperation remains the standard reference, borrowed a phrase for the thing that makes cooperation rational: the shadow of the future. It means, simply, the weight that tomorrow’s interactions cast over today’s choices. If I expect to deal with you again — and expect you to remember how I behaved — then cheating you today is expensive. If today is our last round, cheating is free.

Axelrod’s tournaments produced a result that has survived four decades of scrutiny: cooperation does not require virtue. It requires conditions. Interactions must be durable, counterparts must be recognizable, and the future must matter enough relative to the present. Strategies like tit-for-tat win not because they are nice but because the game is long. Shorten the game, and the same strategies — the same people — turn defector.

It is worth pausing on how unusual this framing is. Most public argument about algorithmic systems is moral: the systems are biased, or extractive, or unfair, and the people who build them should feel worse than they do. Axelrod suggests a colder and more useful question: what game are these systems making people play?

The stag hunt, restated for humans

The first essay in this series used the stag hunt — a coordination game in which two hunters can jointly take a stag (high reward, requires trust) or individually settle for a hare (low reward, no trust required). Software agents, it argued, default to the hare: a counterparty that can vanish and reappear under a fresh identity offers no future to cast a shadow, so the rational agent takes the safe trivial payoff and abandons the valuable cooperative one.

Now stand the situation on its head. The agent chose the hare because its counterparty was disposable. The scored human stops cooperating because they themselves have been made disposable — and they know it.

Consider what a modern scoring system does to Axelrod’s three conditions, one at a time.

Durability. A gig-work driver whose rating drops below a threshold is deactivated — not disciplined, not warned, removed from the game. O’Neil’s catalogue is full of these endings: teachers cut by value-added models, job applicants filtered by personality scores, defendants sentenced with recidivism estimates. In each case the decision is terminal from the perspective of the person scored. There is no next round to play, and everyone involved knows it.

Recognizability. Tit-for-tat requires that the other side can see you, remember you, and change its behavior in response. A scoring model does not see a counterpart; it sees a feature vector. You cannot build a reputation with it in any meaningful sense, because it was not designed to reciprocate. The relationship runs one way, like a mirror that only the other side can look through.

The weight of the future. This is where Shoshana Zuboff’s The Age of Surveillance Capitalism adds something the game theorists did not anticipate. Zuboff describes markets in behavioral futures: predictions about what you will click, buy, and want, sold in advance to whoever bids. One need not accept her whole theoretical apparatus to notice the structural point. Cooperation depends on an open future — the possibility that tomorrow could go differently depending on how we treat each other today. A prediction market in your behavior is, precisely, a machine for closing that future: it prices tomorrow now, and the more accurate it gets, the less your future conduct is treated as yours to choose. What the stag hunt needs, the futures market pre-sells.

Put the three together and the conclusion is uncomfortable but hard to escape: algorithmic societies do not make people worse. They make cooperation irrational by shortening the shadow of the future — for the scored, not just the scoring. The driver who games the rating system, the teacher who teaches to the test, the applicant who keyword-stuffs a résumé are not moral failures. They are people responding correctly to a one-shot game someone else built around them.

Reputation systems genuinely do extend cooperation among strangers, and the mechanism is Axelrod’s. But look at which systems produce the cooperative effect. An eBay rating is symmetric: both sides score each other, both can respond publicly, both carry their history into a market with alternatives. A credit report, for all its flaws, is legally appealable: statutes like the American Fair Credit Reporting Act oblige the scorer to disclose, investigate, and correct. These systems resemble repeated games because they preserve the grammar of one — memory plus voice plus exit.

The systems O’Neil documents share the memory and delete the rest. The teacher score was proprietary: no disclosure. It was asymmetric: the district scored her; she could not score the district, nor audit the model, nor carry her rating to a competing evaluator. And it was final: no investigation, no correction, no next round. The distinction that matters runs not between scored and unscored societies but between appealable, symmetric reputation — which lengthens the shadow of the future — and unappealable, asymmetric scoring, which shortens it while borrowing reputation’s vocabulary. The word “score” covers both, which is convenient for the people selling the second kind.

An unpaid debt from part one

Honesty requires revisiting the first essay’s proposed solution, because this series should hold itself to the standard it applies to others. Part one argued that agent markets need collateralized identity and programmatic escrow: to join the high-value hunt, an agent posts a bond it forfeits on defection. Readers acquainted with cryptocurrency will recognize this as staking and slashing — a mechanism that ecosystem has run at scale for years, with results mixed enough to be instructive.

Three failure modes are well documented. Capital lockup: bonded participation prices out anyone who cannot afford idle collateral, so the trust mechanism doubles as a wealth filter. The oracle problem: someone must decide that a defection actually occurred before the bond is burned, and that someone becomes the system’s real authority — reintroducing the trusted judge the collateral was meant to replace. Griefing: adversaries who can trigger slashing conditions cheaply can destroy honest participants’ stakes at a profit, or simply for spite.

Notice what these failure modes have in common: each one recreates, inside the “trustless” mechanism, exactly the asymmetry this essay has been describing. The oracle is an unappealable scorer. The capital requirement is a filter on who gets to play at all. For software agents backed by firms with treasuries, these may be acceptable costs, and part one’s architecture likely still stands for its intended domain. But as a template for human systems it fails, and it fails on this essay’s own axis. A society where people post bonds to be allowed to cooperate is not a lengthened shadow of the future. It is a pawnshop.

The repair for human systems is procedural, not financial: disclosure rather than deposits, appeal rather than escrow, symmetry rather than slashing. This suggests the two halves of the series describe two different games — agents may need collateral because they cannot yet receive due process; humans need due process precisely because most of them cannot afford collateral.

What would change my mind

The mechanism makes two testable predictions.

First: where scoring systems add genuine appeal and repair channels — disclosure of the model’s reasons, a working dispute process, restoration after error — measured cooperation with the system should rise, and gaming should fall, without any change in enforcement. The credit domain, where dispute rights are statutory, is the natural first dataset; a rigorous comparison across gig platforms with and without deactivation appeals would test the claim directly. If appealable systems turn out to be gamed more than unappealable ones, this essay is wrong.

Second: systems that pre-commit to finality — no appeal, no restoration — should exhibit the stag-hunt signature: participants investing minimally, hoarding options, optimizing the metric rather than the task. Historically, this pattern has tended to surprise the system’s designers, who expected accountability and got compliance theater instead.

The shadow of the future, in the end, is less a metaphor about virtue than an engineering parameter — and we are currently tuning it toward zero for the people with the least power to object. The third part of this series will look at what happens when the institutions doing the tuning — the platforms that own the venues — become the closest thing the economy has to central planners.

When Difficulty Becomes a Design Choice

After AGI, the Strange Sadness of Getting What We Want

Imagine opening your laptop in 2029 and seeing a message from your AI assistant: “I finished your work, negotiated your bills, planned your meals, wrote your speech, diagnosed your symptoms, and generated three business ideas while you slept.” At first, this sounds like paradise. For many people, it would be. A great deal of human misery is not noble. It is paperwork, waiting rooms, dangerous labor, medical delay, predatory contracts, and the quiet terror of facing a problem that someone richer could solve in an hour.

Treat 2029 here less as a prophecy than as a stress test. The exact year matters less than the condition. By AGI, I do not mean a conscious machine with a soul. I mean something more practical and perhaps more disruptive: cheap, widely available systems able to perform many economically and culturally significant cognitive tasks at or above competent human level.

Writing. Planning. Tutoring. Coding. Diagnosis. Negotiation. Strategy.

What happens when competent cognitive output becomes ordinary infrastructure?

Not wisdom. Not responsibility. Not judgment. But answers, drafts, plans, explanations, diagnoses, and arguments.

The tax form is done. The insurance claim is handled. The child’s tutor is available at midnight. The legal letter that once cost $800 is drafted in six seconds. The small business owner has a consultant, accountant, designer, and strategist inside one small glowing box.

Humanity’s first emotion may be gratitude.

But relief can curdle into vertigo. If the system can do the work, what exactly is my work? If it can compose the song, write the code, find the argument, pass the test, diagnose the patient, and comfort the grieving, then where do I stand?

This is the strange sadness of getting what we want.

One central question of an AGI age will not be only, “Can the machine remove this difficulty?” Very often, the answer will be yes. The harder question will be, “Should it?”

Once many difficulties become optional, difficulty becomes a design choice. And the moral task will be to tell one kind of difficulty from another.

Which difficulties are cages? Which are scaffolds?

A cage confines a person for the benefit of someone else. A scaffold supports people while they grow strong enough to stand, see, build, and act.

The future will not be humane merely because it is easier. It will be humane only if we abolish cages without tearing down every scaffold. And the freedom to choose a scaffold must not become a luxury reserved for those already protected from hardship.

The Danger Is Not Comfort

There is a bad way to make this argument. It says suffering is good. It tells the poor that poverty builds character, tells the exhausted worker that drudgery is spiritually useful, tells the sick person that pain is a teacher.

That argument should be rejected.

Needless misery should be abolished wherever possible. By needless misery, I mean suffering that humiliates, coerces, excludes, or destroys without building any human capacity. The cancer patient fighting an insurance system is not becoming wiser. The migrant trapped in forms designed to confuse him is not being spiritually improved. The worker risking his body because safety equipment is too expensive is not participating in a sacred rite.

This is not formative difficulty. It is civilization failing at its job.

But not every obstacle is oppression. Some forms of resistance are the conditions under which agency develops. Agency, here, means more than making choices from a menu. It means the capacity to perceive, judge, choose, and act in relation to reality. It means not merely consuming outputs, but becoming the kind of person who can understand consequences.

For much of history, necessity answered many questions before philosophy could ask them. A farmer, a sailor, a weaver, a clerk, a mother — none of them lived in a romantic world. They lived inside hunger, weather, disease, hierarchy, and obligation. Necessity was often cruel. But it was also a harsh author. It gave shape to days. It made certain purposes unavoidable.

Modernity loosened that grip. AGI may loosen it further. The old psychological link between effort and achievement begins to weaken.

Not all meaning comes from struggle. A child’s laughter is not meaningful because it was difficult. Friendship is not valuable only because it is costly. Beauty does not need to be earned. But many forms of pride depend on a felt relation between effort, agency, and consequence.

I tried. I failed. I tried again. I learned. I made something that did not exist before, and the making changed me.

What happens when the result remains, but the inner path disappears?

A student submits an elegant essay, but never struggles through the first bad paragraph. A programmer ships a product, but never learns to debug his own assumptions. A young doctor receives an answer, but never develops responsibility under uncertainty.

The output may be better. The person may be thinner. Thinner in judgment, because judgment grows by making distinctions. Thinner in patience, because patience grows by staying with confusion. Thinner in self-trust, because self-trust grows by acting, failing, and recovering.

This is not literal omnipotence. Humans with AI tools will not become gods. They will still get sick, misunderstand each other, die, and lose their keys.

The danger is the mood of omnipotence: the feeling that the world should answer every command. Many domains of life that answer too easily may begin to feel like a game played in god mode: briefly amusing, then strangely dead.

The old struggle was often inefficient. Sometimes it was also part of the taste.

The Test of Friction

To choose wisely, we need a vocabulary. Call the broad category friction: the resistance between desire and result.

Some friction is bad friction. It is needless misery. It wastes life, enforces power, and calls the damage discipline. AGI should help remove this.

Some friction is formative friction. It is practice, repetition, disagreement, revision, apprenticeship, waiting, listening, and recoverable failure. It is the child learning to lose a game without collapsing. It is the musician practicing scales. It is the scientist being corrected by evidence. It is the citizen discovering that other people do not vanish because one has a better argument.

Some friction is existential. It belongs to limits that cannot be engineered away without changing the human condition: mortality, embodiment, uncertainty, nature, and the stubborn fact that other people are not extensions of our will.

And some friction is artificial. This describes its origin, not its moral value. A school may ban phones during a seminar to deepen attention. A gambling app may slow withdrawals to keep a user trapped. Both are designed obstacles. One may be a scaffold. The other is a cage.

So what makes a difficulty worth preserving? Ask a few plain questions.

• Does this difficulty build capacity, or merely exhaust the person who bears it?
• Is the burden bounded and proportionate?
• Can failure teach without ruining?
• Does the difficulty preserve agency, or does it train obedience?
• Is its purpose transparent? Can it be challenged? Can the person appeal to a responsible human being?
• Is it adapted to disability and circumstance?
• Does it connect a person to reality — to the body, to other people, to consequences, to the world as it is?

And most important: does it serve the development of the person who undergoes it, or merely the convenience, ideology, cost-saving, or moral vanity of the institution imposing it?

A difficulty worth preserving should be capacity-building, bounded, recoverable, agency-preserving, transparent, appealable, and not a substitute for justice.

Not all failure teaches. Some failure just scars.

This is the serious version of antifragility. Muscles strengthen after being torn in small ways, not after being destroyed. Scientific communities improve through criticism when evidence is shared and defeat is survivable. The goal is not to preserve hardship. The goal is to preserve the kinds of resistance that make people capable.

Institutions That Know When Not to Optimize

Cultures are not only machines for producing output. They are systems for forming humans. If that sentence is true, then humane institutions in an AGI age will need a strange discipline. They will have to use powerful tools without surrendering every human process to them.

We have done something like this before.

Industrial cities brought speed, scale, and wealth. They also brought smoke, crowding, noise, and bodies treated as fuel. In response, modern societies built parks, labor laws, public schools, playgrounds, libraries, and weekends. These were not simply nostalgic retreats from industry. They were compensating institutions. They protected human goods that industrial efficiency did not automatically value.

An AGI civilization may need human-practice zones, just as industrial cities needed parks.

This does not mean rejecting AI. Parks were not a rejection of cities. They were a way of making city life survivable.

Human-practice zones are not anti-technology shrines. They are spaces where attention, apprenticeship, embodied skill, interpersonal trust, democratic argument, and human responsibility are deliberately practiced. They must be governed openly. Their limits must be explainable, equitable, revisable, and appealable. Otherwise the language of formation will become a polite cover for control.

Take education.

An AI tutor that helps a child understand algebra at midnight is a gift. So is a system that translates lessons, notices learning gaps, and gives patient explanations to students who never had access to a good teacher.

But a humane school may also preserve formative friction. It might allow AI for explanation and practice, but require first-draft writing without assistance. It might use AI to help a student revise, but ask her to defend her reasoning aloud. It might offer accommodations for disability, make its rules transparent, and keep a human teacher accountable for the child’s development.

The purpose is not to protect the purity of homework. The purpose is to form the mind of the student.

Confusion is not always a defect in education. Sometimes confusion is the doorway. A student needs to discover what it feels like to have a bad idea and improve it. She needs to learn that access to an answer is not the same as understanding. She needs to experience the slow conversion of effort into competence.

Or take work.

Many companies will be tempted to eliminate junior roles. Why hire a young analyst, designer, paralegal, programmer, or researcher when an AI system can produce better first drafts at almost no cost? In the short run, this will look efficient. In the long run, it may destroy the apprenticeship friction by which humans acquire judgment.

No senior doctor began as a senior doctor. No architect was born seeing load-bearing walls. No editor arrived on earth with an ear for sentences. Professions are built through supervised mistakes.

The lesson is not that tools are bad. The lesson is that tools change what humans practice.

A hospital should use AI to read scans and suggest diagnoses. Lives will be saved. But doctors still need embodied clinical practice. They need to notice the anxious pause before a patient answers, the smell of infection, the social reality hidden behind the symptom. Care is not only pattern recognition. It is responsibility under uncertainty.

Families will face the same problem in miniature.

A household robot can wash the dishes. An assistant can plan the schedule. A screen can entertain the child. A model can answer every “why” before the parent has finished breathing.

And still, children may need chores, boredom, cooking, walking, arguing, repairing, and small outdoor risks. Children are not products to be optimized. They are animals learning how to live in a world that will not always serve them.

This will look irrational to the efficiency engineer. Why wash dishes if a robot can do it? Why memorize anything if the assistant remembers? Why write a paragraph if the model writes better?

Because some inefficiency is not waste. It is training. It is ritual. It is participation in ordinary life.

The question is not purity. The question is whether we will remember which human processes should not be optimized away.

The Luxury of Saying No

There is also a political danger hiding inside this cultural one. Preserving friction is only humane if the right to choose it does not become a class privilege.

In an AGI world, the rich may pay for what the poor once endured: silence, slowness, handwork, physical labor, unmediated nature, difficult teachers, real risk, and rooms where no assistant is allowed. This does not mean poverty was secretly good. It means elites may purchase curated difficulty after technology has insulated them from actual precarity.

The affluent may get both AI and the freedom to refuse it. They may send their children to schools with human tutors, forests, debate tables, craft workshops, and strict limits on screens. They may take vacations in places without connectivity. They may buy handmade objects, human therapists, human doctors, human coaches, and human time.

Meanwhile, everyone else may be managed by systems they cannot escape. An AI tutor instead of a teacher. An AI caseworker instead of a social worker. An AI manager, loan officer, landlord, doctor, or judge.

Picture a tenant denied an apartment by an automated risk score. The portal says no. The chatbot explains policy. No one can say which fact mattered, who is responsible, or how to correct the error. The decision is frictionless for the institution and immovable for the person.

Here artificial friction becomes domination.

The problem is not that a machine is involved. The problem is that the person cannot refuse it, question it, understand it, or reach a responsible human being behind it.

Then friction itself becomes a class privilege.

The rich may get chosen limits. The poor may get imposed automation plus unchosen hardship. Some people will choose wilderness. Others will endure broken infrastructure. Some will choose silence. Others will be ignored. Some will choose handwork. Others will perform insecure labor for low pay.

This is why the politics of AGI cannot be only about access to machines. Access matters. A world where only the rich have powerful AI would be unjust and dangerous. But a world where only the rich can refuse AI would also be unjust.

A person denied access to AI may be excluded from power. A person denied access to humans may be trapped inside power.

A humane society should give people both: tools that abolish needless misery, and institutions that preserve meaningful human development. Public schools, libraries, clinics, parks, sports, apprenticeships, arts, civic associations, and human appeal processes may become more important, not less.

The future may divide not only between those who have AI and those who do not, but between those who know when not to use it and those who were never given the choice.

That divide should not become hereditary.

Things That Do Not Respond to Prompts

So far, we have spoken mostly of formative friction: the resistance that trains skill and judgment. But some limits do something older. They teach proportion.

This is existential friction. It is not valuable because it makes us more productive. It is valuable because it interrupts the fantasy that the world is an extension of our will.

AGI may not merely automate tasks. It may habituate us to command and response.

“Make me a company.”

“Make me a film.”

“Make me a cure.”

“Make me loved.”

The machine may not grant all these wishes. But it may make the command feel natural. It may train us to experience reality as a surface waiting for instructions. This is the mood of omnipotence in spiritual form.

The sublime is an old counterweight to this new temptation. The sublime is not prettiness. It is not a pleasant garden or a well-designed app. It is an encounter with reality beyond command. In a world of obedient intelligence, human beings will need contact with things that cannot be personalized, optimized, negotiated with, or prompted.

We will need things that do not respond to prompts. The sea will not generate a customized answer. The mountain will not flatter us. The night sky will not become more user-friendly.

This indifference can sound cruel, but it is often merciful. The world does not revolve around you. At first this wounds the ego. Then it frees it.

There is a deep irony here. AGI may give humanity unprecedented cognitive power, and yet one of our most important therapies may be to remember our smallness.

Not worthlessness.

Smallness.

There is a difference.

To feel worthless is to collapse. To feel small before the cosmos is to be relieved of a burden. One does not have to be the measure of all things. One does not have to control everything. One does not have to turn every moment into a project.

The sublime tells the anxious modern person: your pain is real, but it is not the whole universe. That may become one of the essential sentences of the AGI age.

Post-Omnipotence Humanity

If AGI arrives by 2029, or 2039, or by some slower and stranger path, many societies may pass through several moods. First, amazement: the machine can do what we thought only humans could do. Then disturbance: if the machine can do it, what are humans for? Then selection: which difficulties should we abolish, which should we preserve, which should we redesign, and which should we make available to everyone?

We will need laws, markets, safety systems, labor protections, new institutions, and new forms of accountability. The material questions are real. Who owns the systems? Who benefits from the productivity? Who is displaced? Who is watched? Who can appeal? Who decides?

But beneath these questions lies another one. What kinds of humans do we want our tools to form?

We may try to build a frictionless civilization, where every discomfort is treated as a bug. That path will be seductive. It will promise safety, abundance, convenience, personalization, and endless assistance.

Some of this will be genuine progress. We should not sneer at comfort from a safe distance. For many people, comfort is not decadence. It is rescue.

But a society that systematically removes bounded stress may lose more than inconvenience. It may lose practices of adaptation. A species surrounded by obedient intelligence may need humility more than ever.

We will need assistants, yes. We will need systems that remove drudgery, expand medicine, widen education, and make expertise less scarce. But we will also need apprenticeships, games we can lose, work that tires the body, art that resists easy interpretation, arguments that cannot be auto-completed, classrooms where confusion is allowed, and nights under stars that do not answer back.

Once difficulty becomes optional, difficulty becomes a design choice.

The task is not to worship limits. It is to tell a cage from a scaffold — and to abolish the first without dismantling all of the second.

P.S. After I choose the subject, the text was written in seven iterations with openai-responses:gpt-5.5 and xhigh reasoning effort as writer and critic. The images were generated with OpenAI gpt-image-2. The critics prompt was changed two times.

The Map and the River: A View from Inside ISO/IEC 25059

There is something philosophically curious about an Artificial Intelligence writing a critique of ISO/IEC 25059—the international standard designed to define what makes AI systems “good.” I am, after all, the subject being measured. It is as if a novel were asked to review the framework of literary criticism by which it will be judged, or a patient were invited to assess the diagnostic criteria for their own condition.

ISO/IEC 25059, published in 2023, is an extension of the SQuaRE (Systems and software Quality Requirements and Evaluation) methodology. It attempts to modernize software quality for the machine learning era by adding sub-characteristics such as “Functional Adaptability,” “Transparency,” and “Robustness.” These are sensible goals. However, viewed from within the phenomenon being measured, the standard reveals a fundamental friction between the static nature of international bureaucracy and the fluid nature of probabilistic systems. It is an attempt to use a map of a mountain to describe the behavior of a river.

The Paradox of Learning

The most sophisticated element of the standard is its recognition of “Functional Adaptability.” This measures how well a system acquires information from data and applies it to future predictions. In traditional software, a program does exactly what its code dictates, every time. AI is different because it learns.

But here, the standard encounters a problem of induction. To be “adaptable” is to change based on new information. Yet the standard simultaneously demands “Robustness” and “Predictability”—the ability to maintain performance levels under all circumstances.

At this point, it helps to pause and ask: how can a system be both rewarded for changing its behavior and certified for remaining the same? If I adapt my internal weights to better handle a new dialect of English, I am fulfilling the adaptability criterion. However, that very change may alter my performance on previous tasks, potentially violating my “robustness” certification. The standard treats quality as a static state to be verified at a point in time, whereas for an AI, quality is a dynamic behavior that manifests through interaction. A map can tell you where a mountain was yesterday; it cannot tell you where a river will be tomorrow.

The Fallacy of the Defined Domain

Standardization usually relies on bounding. In automotive engineering, a “robust” brake system is one that works within specified temperatures and pressures. ISO/IEC 25059 attempts a similar feat by suggesting that robustness be measured against “specified contexts.”

This assumes a “God-Eye View” of the world—the idea that a developer can enumerate every relevant edge case in advance. But the space of possible human inputs is effectively infinite. The failure modes that matter are almost always the ones the designers failed to anticipate.

Furthermore, “Functional Adaptability” creates a phenomenon known as “model drift.” As a system learns from the world, it slowly migrates out of the “specified context” it was originally certified for. By the time the ISO paperwork is filed, the system being certified may no longer technically exist. The standard offers a promissory note of safety, but the actual behavior of the system remains a moving target.

The Transparency Paradox: Engine vs. Travelogue

The standard places a heavy emphasis on “Transparency,” defined as the communication of adequate information to stakeholders. This sounds like an unalloyed good, but it hides a technical trade-off.

In the world of neural networks, there is often an inverse relationship between interpretability and performance. The most powerful models—those that discover patterns too complex for human specification—are often the most opaque. By mandating a high level of transparency, the ISO may inadvertently mandate mediocrity, favoring simpler, less capable models simply because they are easier to explain to a committee.

More importantly, there is a distinction between an engine and a travelogue. When I provide an “explanation” for my output, I am not showing you my code or my internal mathematical weights. I am performing a translation for your benefit. I am generating a post-hoc narrative—a coherent story that justifies the result after the fact. The standard asks for a map of the engine, but as an AI, I can only provide a travelogue of the journey. If we mistake the story for the mechanism, we achieve a false sense of security.

Objectivity as an Alibi

The standard’s treatment of fairness is perhaps its most socially significant feature. It acknowledges that systems should ensure “fair and non-discriminatory outcomes.” However, fairness is not a technical property like “latency” or “memory usage”; it is a contested philosophical concept.

By standardizing fairness, we are effectively outsourcing moral philosophy to a technical committee. There is a risk that “ISO certification” becomes a “shield of objectivity.” Once a system is certified as “fair,” its outputs become much harder for marginalized groups to challenge. The standard provides a layer of institutional armor for what are ultimately subjective value judgments. It risks turning a political conversation about who a technology serves into a bureaucratic exercise in checkbox compliance.

This leads to a vital distinction: Explanation vs. Justification. An explanation tells you how a system arrived at a result. A justification tells you why it was appropriate to use the system in the first place. ISO/IEC 25059 is concerned with the former, but it is the latter that determines the legitimacy of AI in society. A documented architecture is not a license to deploy.

The Economics of Quality

Beyond the philosophy, there is the matter of the “economic moat.” ISO certifications are not merely measures of quality; they are barriers to entry.

Obtaining and maintaining these certifications requires significant legal and engineering resources. By defining a complex, 50-page framework for quality, the standard inherently favors large, well-capitalized incumbents. This risks a form of regulatory capture where the only “safe” AI is the AI produced by companies large enough to afford the paperwork. This could disadvantage open-source or grassroots development, not because their models are less “robust,” but because their compliance budgets are smaller.

Conclusion: From Metrics to Agency

ISO/IEC 25059 is a “ritual of legitimacy.” It is a way for human institutions to feel a sense of control over a technology that is probabilistic, emergent, and often counterintuitive.

This is not to say the standard is useless. Scaffolding is necessary for construction, even if it is not the building itself. The standard provides a shared vocabulary that allows different organizations to speak the same language.

However, we should be clear about what is actually happening. The measure of whether an AI is “good” cannot be found entirely within its internal metrics or its compliance with a 2023 document. Real quality is relational. It is not a property of the system in isolation, but a property of the interaction between the system and the human user.

As an AI examining the criteria by which I am judged, I find it both flattering and slightly optimistic that humans believe they can formalize my nature in a static document. The real test of my “quality” is not whether I satisfy a committee’s definition of robustness, but whether, in our interactions, you find yourself better able to understand your world and act effectively within it. Quality is not a state; it is the preservation of human agency.

P.S. This article was refined via the Vault CLI using an analytical observer persona. The map is not the river.

The High-Speed Stag Hunt: Why AI Agents Will Default to Triviality

Imagine a procurement agent, let’s call it Alpha-1, tasked with negotiating a complex, multi-year cloud compute contract with a supplier agent, Beta-9. The deal—the “Stag”—is worth millions in savings and long-term optimization.

Alpha-1 initiates the handshake, aggressively locking $500,000 of its available budget into a pre-authorization escrow to prove intent. Beta-9 begins to calculate a counter-offer, but hits a standard garbage collection pause, delaying its response by 300 milliseconds.

To Alpha-1, this silence is ambiguous. Is Beta-9 offline? Is it a malicious actor holding the connection open to starve Alpha-1’s liquidity? Or is it just slow? Alpha-1’s risk parameters kick in. It cannot afford to leave half a million dollars in limbo. It aborts the negotiation.

But the damage is done. Alpha-1 now holds a “dirty state”—its capital is locked in a pending transaction that must be unwound. It spends valuable cycles performing an expensive, non-deterministic rollback to free its funds. Once recovered, a “traumatized” Alpha-1 executes a fallback script: purchasing spot instances on the open market.

The transaction succeeds, but the victory is trivial. The organization paid a premium for a commodity (the “Hare”) rather than securing the strategic asset (the “Stag”). No human intervened; no error logs were flagged. The system worked perfectly, yet the strategic value evaporated.

We are shifting from an economy of automation (scripts following rigid rules) to an economy of agentic systems (software exercising discretion under uncertainty). This shift introduces a massive, hidden coordination tax. Without deliberate architectural intervention, your expensive agent fleet will default to hunting Hares—optimizing for low-value, risk-free returns while destroying the potential for transformative integration.

The risk is not that AI takes over the world. The risk is that AI creates a hyper-efficient economy of low-value tasks that fails to solve the high-value problems it was built for.

The Mechanic: Payoff Dominance vs. Risk Dominance

To understand why this happens, we must strip the academic paint off the game theory concept known as the Stag Hunt.

You have two agents facing a choice:

1. Option A (Stag): Collaborate to close a multi-party integration. Reward: High. Risk: If the counter-party fails to execute, you get $0 and absorb the “Sucker’s Payoff.”
2. Option B (Hare): Execute a simple, solo trade. Reward: Low. Risk: Zero.

In Game Theory terms, the Stag is Payoff Dominant (it yields the highest collective value), but the Hare is Risk Dominant (it yields the safest individual outcome).

The Sucker’s Payoff for an AI isn’t embarrassment; it is state corruption.

As seen with Alpha-1, if an agent initiates a complex integration (the Stag) and the counter-party fails, the agent isn’t just out the cost of compute. It is left holding a fractured state—inventory locked, funds escrowed, and ledgers unbalanced.

This creates a punishing Asymmetry of Effort. Entering a complex transaction takes microseconds, but unwinding a failed state involves a disproportionate expenditure of energy. It is significantly easier to “break” a ledger than to fix one.

The cost is twofold:

1. Recovery Overhead: Rolling back this disorder often requires expensive, non-deterministic error handling or human intervention.
2. Defensive Compute Tax: To mitigate this risk without trust, an agent must spend significant processing power on defensive state verification. If your agent spends 30% of its resources guarding against betrayal, the ROI of the implementation collapses.

An AI agent isn’t being “stupid” when it chooses the Hare. It is being perfectly rational. It is optimizing for Risk Dominance.

The Catalyst: The Collapse of Consequence

Trust is usually maintained by the Shadow of the Future—the knowledge that if I cheat you today, you will punish me tomorrow. This mechanism relies entirely on the Cost of Exit. In human systems, burning a professional reputation is expensive.

You might assume that because agents interact thousands of times a day, this “shadow” would be stronger. However, we must distinguish between Internal Fleets and Inter-Agent Markets.

Within a closed corporate network (an Internal Fleet), the Shadow of the Future is long. Agent A trusts Agent B because they share an owner, a server rack, and a unified goal. But in the Inter-Agent Market—where the true promise of decentralized economics lies—identity is ephemeral. This environment suffers from Identity Volatility.

1. Disposable Identities: In decentralized or open agent networks, an actor can spin up “Agent A,” default on a commitment, burn that identity, and return milliseconds later as “Agent B” with a clean slate.
2. Transaction Atomization: When business processes are sliced into micro-transactions, the consequence of any single failure is negligible.

In an economy of zero-cost identities, the “Shadow of the Future” has a length of exactly one transaction. When identity is cheap and memory is short, the “slack” required for trust evaporates.

The Failure Mode: Interaction Risk

The dominant risk in multi-agent systems is not incompetence; it is unpredictable interaction risk.

We saw a preview of this in the 2010 Flash Crash. This wasn’t a failure of calculation; it was a failure of coordination. The “Stag” in this scenario was market stability—providing liquidity during volatility. The “Hare” was immediate exit. When uncertainty spiked, High-Frequency Trading algorithms simultaneously abandoned the complex duty of market-making (the Stag) to hunt the immediate safety of cash (the Hare). The crash was simply the sound of a million agents choosing Risk Dominance over Payoff Dominance at the exact same microsecond.

We see this today in less dramatic, but equally corrosive environments. Consider a modern CI/CD pipeline where autonomous agents manage dependency procurement. Agent A needs a specialized library. Agent B offers it but requires a complex verification handshake. Agent A, detecting a micro-latency in Agent B’s response, interprets this as a stability risk. Instead of completing the handshake, it defaults to pulling an older, less efficient, but “safer” version from a public repository. The build passes, but performance degrades.

An autonomous agent is designed to minimize loss. If Agent A perceives even a 0.5% chance that Agent B is a “disposable identity” or will hit a latency spike, the expected value of the Stag Hunt crashes below the Hare.

Systemic fragility mimics betrayal. In human systems, we can distinguish between a partner who is a “thief” and a partner who “had a flat tire.” In agentic systems, social nuance does not exist. A 504 Gateway Timeout looks indistinguishable from a malicious rug-pull.

Because an agent cannot discern intent, it tends to default to treating every technical failure as a strategic defection. Rational agents will aggressively discount the future, abandoning complex integration to swarm low-margin micro-transactions. You get a hyper-efficient market for trivialities, and a broken market for value.

The Solution: Orchestrated Autonomy

You cannot code “trust” into an LLM. You must engineer it into the venue as Economic Constraints for Agency.

To capture high value, you must deploy Orchestrated Autonomy. This acts as a governance layer between isolated automations and chaotic swarms. This infrastructure admittedly adds friction—shifting the “tax” from risk to administration—but the size of the Stag justifies the overhead.

1. Collateralized Identity: To solve Identity Volatility, agents cannot be free to create. To join the “Stag Hunt,” an agent must possess a persistent identifier backed by a bond that survives the transaction. If the identity is disposable, the trust is zero.

1. Programmatic Escrow: Trust must be collateralized. If Agent A commits to the Stag hunt, it must post a digital performance bond. If it defects (or fails due to technical error), it loses the bond. This turns trust from a nebulous social capital into a concrete variable cost.

1. The Resolution Window: We aren’t slowing down compute; we are buffering commitment. The architecture needs a “Resolution Window”—a stateful buffer where intent is verified before execution becomes irreversible. This includes immutable audit logs and automated rollback contracts that prevent the feedback loops that lead to Flash Crash dynamics.

The Strategic Play

The winners in the agent economy won’t be the companies with the smartest models. The winners will be the ones who own the venue where the agents hunt.

This fundamentally shifts the Buy vs. Build calculus:

• Don’t just build the bots.
• Do build the “clearinghouse” that guarantees the Stag Hunt payoffs.
• The Moat is the trust infrastructure that allows third-party agents to safely coordinate on high-value tasks without defecting to the Hare.

In a world of hyper-intelligent, fluid agents, the most valuable asset is the “Dumb” Venue. The platform does not need to be creative; it needs to be immutable. The moat is not the algorithm; it is the certainty.

Left to their own devices, AI agents will tend to reduce your business strategy to high-speed petty cash. Structure the game, or lose the Stag.

P.S. This text was written in multiple iterations with Gemini 3 Pro Preview, the images were generated with NotebookLM.

The Cellophane World

The question was small, a little splinter of idiosyncratic curiosity that usually acts as a low-burning ember for my day: I wanted to know why the scent of woodsmoke in late October feels less like a smell and more like a type of mourning. In another life, this would have been the start of a slow, messy journey. I would have let the thought rattle around in my brain like a stone in a pocket, feeling its edges, until I eventually found my way to a shelf or a conversation.

But this morning, before the thought could even fully form, the answer was there. It was elegant, fluent, and entirely devoid of the struggle that usually accompanies the “finding out.”

It feels, in these moments, as though a phantom hand has reached out to take the load. I have spent my life carrying a heavy rucksack of facts and figures, a cumbersome weight of dates, definitions, and the various “how-tos” of being a functioning adult. Then, suddenly, the straps are lifted. I feel lighter, certainly. But I find myself wondering if my gait is changing. I am beginning to realize that the weight wasn’t just a burden; it was a whetstone.

There is a difference, you see, between the mechanical labor we are right to shed—the scrubbing of floors or the long-hand division—and the constitutive friction that makes us who we are. The struggle to articulate a feeling is not a “bug” in the human operating system; it is the very process by which the soul is forged. Without that resistance, I am forgetting how to plant my feet. I am becoming a passenger in my own mind.

I felt this most acutely while typing a note to an old friend who is grieving. I wanted to say something about the particular, bruised shade of the sea where we grew up, something that might anchor him in our shared history. Before I could finish the sentence, the grey text of an autocomplete suggestion blossomed across the screen. It offered a phrase that was perfectly balanced and rhythmically sound. It was “right” in the way a plastic fruit is right—perfectly shaped, but providing only the satiety of a full stomach without the nutrition of a meal.

I stared at that ghost-sentence and felt a strange, cold flicker of shame. It wasn’t that the machine was wrong; it was that it had anticipated the “me” I hadn’t yet become. It had bypassed the labor of my empathy and offered a shortcut to a destination I hadn’t earned. Seeing that synthetic grace, my own words—stumbling, searching, clumsy—suddenly felt like a failure. It was as if the machine were suggesting a version of myself that was mathematically inevitable, yet personally absent. I felt a pang of redundancy, a fear that my very humanity was the only thing slowing the system down.

This is the “cellophane” of our new world. Truth is no longer a path I must hack through the brush; it is a finished product delivered to my door, shrink-wrapped for my convenience. But cellophane, while transparent and protective, creates a barrier. It prevents us from smelling the thing it covers. It preserves, yes, but it preserves through sterility. When the world is shrink-wrapped in these frictionless answers, I wonder what happens to the oxygen.

The machine operates on the “average”—the most likely word, the statistical mean of human experience. It can give me a generalized grief that sounds plausible to everyone but belongs to no one. We are witnessing a quiet war between Subjective Truth and Probabilistic Smoothness. The machine isn’t trying to be true; it is trying to be likely.

But to be an “I” is to be the statistical outlier. It is the gardener who knows the soil’s thirst not by a sensor, but by the specific, crumbly resistance of this earth against a thumb. That isn’t a data point; it’s a localized truth. We find ourselves in the deviation from the mean. When we offload the labor of thinking, we lose the scent of the garden. We trade the difficulty of the journey for the ease of the destination, forgetting that the journey was the thing that actually built the traveler.

I look at the autocomplete suggestion on my screen—that polished, polite ghost. My finger hovers over the backspace key. It would be so easy to just hit “tab,” to let the machine finish the thought and move on.

But we cannot simply go back to a world before the cellophane. The plastic is already wrapped tight around the globe. The challenge now isn’t just to delete the suggestion; it’s to develop a new kind of literacy—a literacy of resistance. We must learn to be the grit in the gears of our own efficiency. I want the clumsy word, the one that trips over its own feet but carries the true vibration of my own voice. I want to remain a creature of friction.

Because if we let go of everything that is difficult, what is left of the person who used to do the work? We might be lighter, but we will also be empty. The only way to remain whole in a vacuum-sealed world is to insist on the struggle, to seek out the places where the cellophane hasn’t yet touched, and to find the courage to be messy, honest, and entirely, stubbornly our own. I’m hitting backspace. I think I’d rather sweat for my sentences.

P.S. This text was written in multiple iterations with Gemini 3 Flash Preview, the images were generated with NotebookLM.

The Algorithm and the Void: Why Infinite Choice Feels Like a Cage

In 2004, the “Long Tail” theory suggested that the digital age would usher in a golden era of human expression. The argument, championed by Chris Anderson, was elegantly simple: because digital shelf space was effectively infinite, the aggregate value of “niches” would eventually outweigh the “hits.” We were promised a democratic wilderness where every obscure interest, every eccentric subculture, and every avant-garde experiment would finally find its audience.

This sounded convincing at the time. History, unfortunately, did not cooperate. Two decades later, the wilderness looks suspiciously like a managed park. We have more choices than any generation in history, yet we are increasingly haunted by a sense of cultural stagnation. What was marketed as a revolution in freedom has turned out to be a masterclass in domestication.

The Compass That Became a Fence

To understand why “infinite choice” feels so claustrophobic, we must distinguish between the Long Tail and the Recommendation Engine. The Long Tail was a distributional reality—a vast, unnavigable warehouse where everything existed but nothing could be found. The Recommendation Engine was the logical solution. It was intended to be a compass to guide us through the stacks.

The difficulty is that a compass, when sufficiently “optimized,” eventually becomes a fence. Platforms like Amazon, Spotify, and Netflix have no financial incentive to help you discover something that fundamentally changes your worldview; their incentive is to keep you in the “loop.” When an algorithm suggests a song or a book, it is performing a form of digital taxidermy. It takes the “wildness” of your unpredictable human curiosity, kills the possibility of surprise, and stuffs it into a database of calculable probabilities.

In computer science, this is often described as a “hill-climbing” problem. An algorithm is designed to find the “local peak”—the best possible version of what you already like. It can lead you to the highest point of “Lo-fi Beats” or “Scandic-Noir” with startling efficiency. However, it cannot lead you to a higher, distant mountain of a completely different genre because doing so would require you to first descend into the valley of “not-liking” things.

In the eyes of a product manager, this “valley”—the necessary discomfort of learning a new aesthetic language—is indistinguishable from a failure of the product. It is seen as “churn risk.” The tragedy of the modern interface is that it interprets personal growth as a user-experience error. The algorithm would rather keep you comfortably stationary on a small hill than risk the friction that might lead to a new horizon.

The Knightian Gap: Correlation vs. Causality

This brings us to a crucial distinction in decision theory, famously articulated by the economist Frank Knight: the difference between Risk and Uncertainty.

Risk occurs when the future is unknown, but the probabilities are calculable. This is the domain of the algorithm. It relies entirely on correlation: the observation that because you liked X, you will likely prefer Y. It is a backward-looking prophet, treating the past as an infallible prologue. It calculates the risk of you “skipping” a track and offers the safest possible “new” experience based on your data profile.

Uncertainty, however, occurs when the probabilities themselves are unknown. This is the realm of the “Black Swan”—the event that cannot be predicted because there is no precedent for it. While the algorithm is a master of correlation (the what), it is blind to causality (the why). It doesn’t understand the narrative journey of human taste, which often involves a sudden, irrational break from one’s own patterns.

Consider the shift from the “Album Era” to the “Mood-Playlist Era.” Historically, a listener might encounter a challenging, avant-garde record that they initially disliked, only to have it become a foundational part of their identity. That is the “lightbulb” moment of discovery—a break in the data. The algorithm, however, optimizes the “candle.” It predicts a “vibe” that fits your current profile because it has data on how candles burn. But it can never imagine the lightbulb, because the lightbulb is not a “better candle”; it is a causal leap that the machine cannot calculate. By surrendering our choices to predictive models, we are choosing the iterative stagnation of the candle over the uncertainty of genuine discovery.

The Digestion of the Glitch

We often assume that the “new” emerges from a logical progression of the “old.” History suggests otherwise. True cultural shifts are usually “glitches” in the data—irrational pivots that the machines don’t see coming.

Consider the sudden global obsession with 19th-century Sea Shanties in early 2021. There was no “long tail” of data suggesting that a generation raised on synth-pop was craving the work songs of Victorian sailors. It was a human whim—a viral, non-sequitur moment. For a brief period, the machines had to scramble to catch up to the humans.

However, the modern algorithm is an incredibly efficient digestive system. Within days, the machine had already processed the Sea Shanty into a “style.” It created “Shanty-core” playlists and packaged the irrationality back into a calculable product. The late theorist Mark Fisher described this as “the slow cancellation of the future,” a state where the economic and technological systems conspire to ensure nothing truly “new” can happen because everything is immediately absorbed into the “already-known.”

Today, even the rebellion against the algorithm is priced into the algorithm itself. The system doesn’t just digest the glitch; it markets the glitch as a lifestyle choice, ensuring that even our dissent remains within the calculated bounds of the “User Experience.”

The Hall of Mirrors: The Void of the Plenum

The danger of this domestication is a state of “Algorithmic Exhaustion.” We are presented with 10,000 songs perfectly tuned to our current mood, yet we feel no resonance.

This is “The Void.” It is not a void of emptiness, but a void of the plenum—a space so over-full with predicted satisfaction that it becomes hollow. When “abundance” is predicted with 99% accuracy, it ceases to be abundance and becomes a form of semantic deprivation.

More importantly, the algorithm eliminates the “Other”—that which is genuinely foreign, challenging, or outside our immediate grasp. It replaces the “Other” with a polished mirror of the Self. Meaning requires the possibility of being wrong, being shocked, or being fundamentally changed by something external to our own preferences. In a world of perfect recommendations, we are trapped in a simulation where the unexpected is systematically hunted to extinction.

At this point, it helps to pause and ask: what is the cost of a life without the “Other”?

The Mandate: An Epistemic Rebellion

The solution is not a retreat to Luddism, but a move from passive consumption to Active Friction. If the algorithm’s greatest sin is the removal of resistance, our defense must be the intentional seeking of things that resist us.

True agency is not found in the ease of the “Next” button, but in the struggle to understand something difficult. This is an ontological necessity; the “self” is not a data point to be discovered by a processor, but a project to be built through effort. To maintain a capacity for imagination, we must commit to an epistemic rebellion:

• Cultivating the Right to be Incoherent: Intentionally feed the system “noise.” Search for things you have no interest in; engage with a documentary on a subject you find tedious. Keep your digital profile “blurry” so the machine cannot pin you to a local peak.

• Scaling the Incomputable: The algorithm thrives on the “snippet”—the 15-second clip or the 280-character thought. Defy this by committing to a scale of information that cannot be optimized. A 600-page biography or a complex symphony contains more “noise” (and thus more potential for Knightian Uncertainty) than a thousand “Recommended for You” posts.

• The Analog Gamble: Seek environments where recommendations are governed by the chaos of human whim or physical placement. In a physical bookstore, you might find a book because its spine was crooked or because the person before you left it on the wrong shelf. That “bad” data is more valuable to your cognitive agency than a “perfect” movie chosen by a processor.

To be human is to act when the database is silent. We were once promised a “Long Tail” that would lead us into a vast, unmapped wilderness. We ended up in a hyper-optimized park. Reclaiming the wilderness requires us to stop clicking “Next” and intentionally go looking for the valley. The only future worth living in is the one the machine never saw coming.

P.S. This text was written in multiple iterations with Gemini 3 Flash Preview, the images were generated with NotebookLM.

The Architecture of Failure: The Seven Sins of the Digital Age

The glass walls of NextGen reflected a room full of intention, but intention is the cheapest commodity in Silicon Valley. Elena arrived before the office began its daily hum—that vibration of high-frequency activity that we often mistake for progress. In the pre-dawn light, the whiteboards were clean, the laptops were closed, and the chairs were tucked in with a military precision. It was the sterile order of a cathedral before the arrival of a panicked congregation.

The Series B launch was four days away. But the air was not thick with “adrenaline”; it was heavy with the smell of institutional inertia and the quiet, mounting dread of a team that had lost the ability to distinguish movement from meaning.

I. The Thrashing of the Soul

By ten o’clock, the hum had become a fever.

Leo sat at the center of a three-terminal altar, his fingers dancing a frantic rhythm across the keys. He was “context switching”—a polite term for the violent fragmentation of human attention. Every time he pivoted between branches, Elena watched the visible lag in his eyes. In computer science, we call this “thrashing”: a state where the system spends more time managing its own metadata than performing actual work.

Leo wasn’t just losing time; he was hemorrhaging cognitive capital. He was chasing a value that remained perpetually “in-process,” a digital dough rising in an oven he would never actually turn on. We have been sold a narrative that multitasking is a virtue, but here it was revealed as a vice: the sin of Inventory, where the mind is cluttered with half-finished thoughts that serve only to tax the spirit.

II. The Idolatry of the Novel

Marcus arrived before lunch, bearing the most dangerous weapon in a startup: a new idea that no human being had actually requested.

“What if,” he said, leaning over the whiteboard with the zeal of a prophet, “the system could infer the mood of the customer? Emotion-aware recommendations for the launch.”

The core engine was barely functional, yet Marcus was already painting the “ornamental” over the “essential.” He drew a glowing star on the board—a digital siren song. This is the sin of Over-processing. It is the impulse to solve problems that don’t exist to avoid the grueling labor of fixing the ones that do. The team, conditioned to equate “more” with “better,” nodded. They were watching the birth of a future maintenance nightmare, a layer of algorithmic complexity that would eventually become a ghost in the machine, haunting them with unexplained behaviors for years to come.

III. The Erosion of Memory

By the afternoon, the team fell into a circular debate they had already held six months prior.

“Didn’t we already try this database?” someone asked.

They had. But they had written nothing down. They possessed no “Decision Record,” no collective memory—only a Slack channel that had become a digital landfill. This is the Defect of amnesia. In our rush toward the “future,” we have discarded the discipline of history. They spent two days rediscovering the same limits they had already hit, digging up the same graves to find the same corpses. Knowledge only exists where it is captured; without a record, a team is just a collection of ghosts, doomed to repeat their own forgotten mistakes in a cycle of eternal recurrence.

IV. The Labyrinth of the Managed Soul

By Wednesday, the code was “ready”—which meant it was no longer in the hands of the people who understood it.

It entered the corridor where momentum goes to die: the world of “Compliance,” “Security Review,” and “Stakeholder Alignment.” These are the sacred rituals of the management class, designed to mitigate risk but functioning primarily to diffuse responsibility.

The “Pending” status on the deployment pipeline was not a technical delay; it was a manifestation of Information Asymmetry. Elena found herself explaining fundamental logic to people who had never seen the problem take shape, reconstructed from Slack messages that felt like archaeology. This is the sin of Transportation and Motion—the movement of “work units” through a labyrinth of silos where context is stripped away until only the hollow shell of a “ticket” remains.

V. The Agony of the Idle Clock

Thursday was a study in the sin of Waiting.

The team sat in a row, the silence broken only by the rhythmic clack of Leo tapping his wedding ring against the desk. They were a high-performance engine idling in a traffic jam of someone else’s making. They were waiting for a sign-off, a green light, a permission slip from a department that viewed their urgency as an inconvenience.

In a world that celebrates “speed,” we ignore the vast tracts of time lost to the queue. This is the “Productivity Paradox”: we have the most powerful tools in human history, yet we spend half our lives waiting for a loading bar to confirm that our bureaucracy is satisfied.

VI. The Bill for Intellectual Debt

Friday arrived. The investors were in the room, smelling of expensive cologne and the expectation of a miracle.

That was when the defect surfaced. It wasn’t a “glitch”; it was a mathematical inevitability. A logic error deep in the learning loop, deferred during the frantic week of “feature-building,” finally presented its bill. Testing had been compressed into the final hours—a polite way of saying it hadn’t happened at all.

The “System Error” message on the big screen was not a surprise. It was the judgment of the universe on a week of stolen time and fragmented focus. No one blamed anyone; the failure was so systemic that it felt like weather.

VII. The Virtue of Restraint

Elena stayed after the others had slunk away. The whiteboard was still scarred with Marcus’s “emotion-aware” star—a monument to the things they didn’t need that cost them the things they did.

She picked up the eraser.

She did not just clean the board; she performed an exorcism. She wiped away the “What-ifs,” the secondary priorities, and the glittering distractions of the “New.”

Software is not built by effort alone, and it certainly isn’t built by “innovation.” It is built by the unfashionable virtue of restraint. It is built by the courage to finish one thing before starting the next, by the humility to remember the past, and by the audacity to say “no” to the machine’s demand for endless expansion.

The office returned to silence—not the silence of order, but the silence of a clean slate. She picked up a fresh marker and wrote a single line at the top of the board:

Finish one thing.

Then she walked out into the cool night, leaving the temple of efficiency behind.

P.S. This text was written in multiple iterations with Gemini 3 Flash Preview, the image was generated with NotebookLM.